car-detection-bayes/README.md

76 lines
3.9 KiB
Markdown
Raw Normal View History

2018-12-23 12:50:44 +00:00
<img src="https://storage.googleapis.com/ultralytics/logo/logoname1000.png" width="200">
2018-08-26 08:51:39 +00:00
# Introduction
2018-09-01 12:10:53 +00:00
This directory contains software developed by Ultralytics LLC, and **is freely available for redistribution under the GPL-3.0 license**. For more information on Ultralytics projects please visit:
2018-12-23 12:50:44 +00:00
https://www.ultralytics.com.
2018-08-26 08:51:39 +00:00
# Description
2018-09-02 21:43:26 +00:00
The https://github.com/ultralytics/yolov3 repo contains inference and training code for YOLOv3 in PyTorch. Training is done on the COCO dataset by default: https://cocodataset.org/#home. **Credit to Joseph Redmon for YOLO** (https://pjreddie.com/darknet/yolo/) and to **Erik Lindernoren for the PyTorch implementation** this work is based on (https://github.com/eriklindernoren/PyTorch-YOLOv3).
2018-08-26 08:51:39 +00:00
# Requirements
2018-11-27 17:14:48 +00:00
Python 3.7 or later with the following `pip3 install -U -r requirements.txt` packages:
2018-08-26 08:51:39 +00:00
- `numpy`
- `torch`
- `opencv-python`
2018-08-26 09:05:13 +00:00
# Training
2018-08-26 08:51:39 +00:00
2018-12-11 18:54:31 +00:00
**Start Training:** Run `train.py` to begin training after downloading COCO data with `data/get_coco_dataset.sh`. Training runs about 1 hour per COCO epoch on a 1080 Ti.
2018-09-01 16:35:28 +00:00
2018-12-11 18:54:31 +00:00
**Resume Training:** Run `train.py --resume` to resume training from the most recently saved checkpoint `weights/latest.pt`.
2018-09-01 16:35:28 +00:00
2018-12-28 18:17:33 +00:00
Each epoch trains on 120,000 images from the train and validate COCO sets, and tests on 5000 images from the COCO validate set. Default training settings produce loss plots below, with **training speed of 0.6 s/batch on a 1080 Ti (15 epochs/day)** or 0.45 s/batch on a 2080 Ti.
2018-12-28 18:23:35 +00:00
2018-12-11 18:46:15 +00:00
![Alt](https://user-images.githubusercontent.com/26833433/49822374-3b27bf00-fd7d-11e8-9180-f0ac9fe2fdb4.png "coco training loss")
2018-09-01 11:34:05 +00:00
2018-09-01 11:43:07 +00:00
## Image Augmentation
2018-09-01 11:34:05 +00:00
2018-09-01 16:57:18 +00:00
`datasets.py` applies random OpenCV-powered (https://opencv.org/) augmentation to the input images in accordance with the following specifications. Augmentation is applied **only** during training, not during inference. Bounding boxes are automatically tracked and updated with the images. 416 x 416 examples pictured below.
2018-09-01 11:41:34 +00:00
Augmentation | Description
--- | ---
2018-12-28 18:23:35 +00:00
Translation | +/- 10% (vertical and horizontal)
2018-09-01 12:04:42 +00:00
Rotation | +/- 5 degrees
2018-12-28 18:23:35 +00:00
Shear | +/- 2 degrees (vertical and horizontal)
Scale | +/- 10%
2018-09-01 12:10:06 +00:00
Reflection | 50% probability (horizontal-only)
2018-09-01 12:04:42 +00:00
H**S**V Saturation | +/- 50%
HS**V** Intensity | +/- 50%
2018-09-01 11:34:05 +00:00
2018-12-28 18:52:25 +00:00
<img src="https://user-images.githubusercontent.com/26833433/50525037-6cbcbc00-0ad9-11e9-8c38-9fd51af530e0.jpg">
2018-08-26 08:51:39 +00:00
2018-08-26 09:05:13 +00:00
# Inference
2018-12-22 11:49:55 +00:00
Run `detect.py` to apply trained weights to an image and visualize results, such as `zidane.jpg` from the `data/samples` folder, shown here.
2018-12-22 12:05:52 +00:00
2018-12-22 12:20:40 +00:00
**YOLOv3:** `detect.py --cfg cfg/yolov3.cfg --weights weights/yolov3.pt`
2018-12-28 18:23:35 +00:00
<img src="https://user-images.githubusercontent.com/26833433/50524393-b0adc200-0ad5-11e9-9335-4774a1e52374.jpg" width="800">
2018-12-22 12:05:52 +00:00
2018-12-22 12:20:40 +00:00
**YOLOv3-tiny:** `detect.py --cfg cfg/yolov3-tiny.cfg --weights weights/yolov3-tiny.pt`
2018-12-28 09:50:08 +00:00
<img src="https://user-images.githubusercontent.com/26833433/50374155-21427380-05ea-11e9-8d24-f1a4b2bac1ad.jpg" width="800">
2018-09-04 12:36:51 +00:00
2018-12-22 11:49:55 +00:00
# Pretrained Weights
2018-12-22 11:58:59 +00:00
2018-12-22 11:49:55 +00:00
Download official YOLOv3 weights:
**Darknet** format:
- https://pjreddie.com/media/files/yolov3.weights
- https://pjreddie.com/media/files/yolov3-tiny.weights
**PyTorch** format:
2018-12-28 07:48:10 +00:00
- https://drive.google.com/drive/u/0/folders/1uxgUBemJVw9wZsdpboYbzUN4bcRhsuAI
2018-09-01 16:48:53 +00:00
2018-12-11 19:18:05 +00:00
# Validation mAP
2018-08-26 09:05:13 +00:00
2018-12-11 19:18:05 +00:00
Run `test.py` to validate the official YOLOv3 weights `weights/yolov3.weights` against the 5000 validation images. You should obtain a .584 mAP at `--img-size 416`, or .586 at `--img-size 608` using this repo, compared to .579 at 608 x 608 reported in darknet (https://arxiv.org/abs/1804.02767).
2018-09-02 11:09:05 +00:00
2018-12-22 12:09:05 +00:00
Run `test.py --weights weights/latest.pt` to validate against the latest training results. **Default training settings produce a 0.522 mAP at epoch 62.** Hyperparameter settings and loss equation changes affect these results significantly, and additional trade studies may be needed to further improve this.
2018-08-26 08:51:39 +00:00
# Contact
2018-12-28 18:17:33 +00:00
For questions or comments please contact Glenn Jocher at glenn.jocher@ultralytics.com or visit us at https://contact.ultralytics.com.